Feedback

[mrcevic]

I've been looking at the videos in my spare time. First of all, congrats, this is good stuff. My main interest in the work you're doing is about being able to a) automatically tag images, b) create correlation links between those tags. This could be useful to autoclassify images and be able to suggest similar or related images. One possible use for this type of functionalities is to overcome the problems of manually applying taxonomies. For instance, once could imagine feeding a AI system a bunch of highly performant image assets and getting back an analysis of the most common features within those assets.

[gilo-agilo]

@annagusevacode - please take a look. This can be some extension of the current challenge or a new challenge (as the evolution of the current one).

[annamykcode]

@mrcevic Thanks for your feedback Your propositions are great and useful for a lot of clients

Here are my initial thoughts about them:

a) this task is similar to a very common image classification task (classify what kind of objects are there on the image). There are already a lot of the State of the Art (SOTA) solutions with pre-trained models which can be useful. The only thing here we should consider is that there is not any general solution, which would fit all businesses. I mean, each image needs its own objects to be detected, for instance: Ford would need to tag the model of the car, its color, the background, while Christie's would need whether it is a vase, a painting, or something else. So, if we'd like to satisfy the needs of both of these businesses we'll create a tagging tool that would tag cars' models, colors, vases, paintings, etc. But, if let's say RL would like to use this tool, when they need to tag dresses, its' colors, cuts, etc, our tool won't satisfy their requests. This means that to create the accurate tool for each business we need to know their needs and have a lot of already tagged images to be able to train the accurate Neural Network (NN) model.

b) it could be done. Actually, our current Image retrieval project is doing the same way. The model is fed with the image, gets the features of the image, and searches the images in DB with similar images. And this way could be used for the automatic taxonomy. The only thing is that the NN model creates its own vision of features (which may vary from the human vision of features) and therefore the model can combine some features together, which humans would think are completely different. But that's all can be tuned based specifically for each business.

So, the first task is really common and doesn't need any POC, just enough tagged images to train the model which would further auto-tag any future untagged images. The second task is less common and we can try to create POC of automatical taxonomy creation based on our current Image retrieval task.

[mrcevic]

@annagusevacode thanks for your feedback, I will write a challenge based on the second point and submit it as a separate challenge